An Empirical Analysis of Preprocessing Techniques for Short-Term Electricity Demand Forecasting

This paper analyzes the impact of preprocessing techniques – including encoding, scaling, and engineered features – on the performance of short-time electricity demand forecasting based on machine learning models and validates their effectiveness through statistical hypothesis testing. We evaluated the effects of input data encoding (label, one-hot, cyclical), standardization, and engineered features. Through rigorous experiments with multiple model instances per condition, statistical significance was verified via Wilcoxon signed-rank tests. The results demonstrate that preprocessing techniques generally lead to a statistically significant improvement in electricity demand forecasting performance. However, the experimental results confirmed that the degree of effectiveness varies depending on the specific machine learning model employed. This study empirically highlights the importance of input data preprocessing in short-term electricity demand forecasting and provides insights into effective feature handling strategies considering model characteristics.